Chemlistem: chemical named entity recognition using recurrent neural networks.

Area of Science:

• Computational chemistry
• Bioinformatics
• Natural Language Processing

Background:

• Chemical Named Entity Recognition (NER) traditionally relied on Conditional Random Fields (CRF).
• Deep learning (DL) has shown significant success in various Natural Language Processing (NLP) tasks.
• The potential of DL as an alternative to CRFs for chemical NER warrants investigation.

Purpose of the Study:

• To develop and evaluate deep learning-based systems for chemical Named Entity Recognition (NER).
• To compare the performance of different deep learning architectures against traditional CRF methods.
• To explore the effectiveness of character-level and token-level approaches in chemical NER.

Main Methods:

• Developed a DL system using bidirectional Long Short-Term Memory (LSTM) networks with rich per-token features and word embeddings.
• Created a second DL system focusing on character labeling with character embeddings and multiple LSTM layers, omitting tokenization.
• Implemented an ensemble system combining the outputs of the first two DL systems.
• Utilized transfer learning to further enhance model performance.

Main Results:

• The ensemble system achieved a top performance in the BioCreative V.5 competition.
• Subsequent models using transfer learning reached a final F-score of 90.33% on test data.
• Specific performance metrics include 91.47% precision and 89.21% recall.

Conclusions:

• Deep learning models, particularly LSTMs, provide a competitive and effective alternative to CRF-based approaches for chemical NER.
• Character-level and token-level DL strategies can be successfully combined for improved chemical entity recognition.
• The developed DL systems demonstrate state-of-the-art performance in chemical NER tasks.